15.4 Minimizing Fuel Usage in Long-Haul Aircraft with Optimized Flight Planning Using Ensemble Numerical Weather Prediction Models

Thursday, 10 January 2019: 2:15 PM
North 224B (Phoenix Convention Center - West and North Buildings)
Brendon T Sands, Air Force Institute of Technology, Wright-Patterson AFB, OH; and A. J. Geyer

The U.S Air Force’s Air Mobility Command (AMC) is the single largest consumer of aircraft fuel in the Department of Defense (DoD) (Jensen et al. 2014). With aircraft flying around the world 24 hours a day, seven days a week, AMC could save an average of $28M per year if it were able to save as little as $200 worth of fuel per sortie (Heseltine 2008). While many factors effect in-flight fuel efficiency, winds aloft are the most impactful factor during the kinds of long-haul flights frequently flown by AMC aircraft. Therefore, determining beforehand which flight routes minimize fuel consumption (based on winds aloft) is vital during flight planning. Failing to select the most efficient flight routes will result in wasted fuel, which translates into wasted money (Homan 2014). AMC contractors currently use the deterministic Global Forecast System (GFS) Numerical Weather Prediction (NWP) model for aircraft route planning (AMC 2017). Krishnamurti et al. (2000) found that ensemble models possessed more forecast skill than their deterministic counterparts across all of the NWP models inspected. As such, AMC has wished for years to incorporate ensemble NWP into their flight planning processes. A previous effort by Homan (2014) attempted to accomplish this by replacing the GFS model data in the current AMC flight planning process with the mean values of the Global Ensemble Forecast System (GEFS) model data. That research found that for a given route, the GEFS means provided an improved estimate of overall fuel consumption than that provided by the current GFS-based AMC process. However, the Homan (2014) approach was not able to predict which choices of flight routes would provide the minimal fuel usage to fly from one location to another. Boone (2018) took a new approach to incorporating the GEFS into the AMC flight planning process. That research effort used a model from Reiman (2014) to convert U and V component winds from the GEFS into fuel burn calculations for a C-17 cargo aircraft to either fly level, climb or descend between any two points. Since the winds between any two adjacent model points are correlated, using the mean values at each point would increase the resulting error in any estimate of fuel consumption while also increasing the probability of selecting a suboptimal route. Boone (2018) treated each GEFS member as a sample of a network where the model data points are the nodes and the fuel burn estimates are samples of the arc lengths in the network. Since winds are correlated between adjacent points, the adjacent arc lengths in these networks were also correlated. Ensembles generate independent identically distributed (IID) output. Thus, these ensemble member networks constituted an IID sample of a single stochastic network with correlated random arcs where the sample is taken for the entire network at a time. Boone (2018) then used the classic Shortest Path Problem (SPP) one each ensemble member network one at a time to build a candidate list of fuel efficient flight routes. The list of candidate routes was then fed back through every ensemble member to generate an IID sample of fuel burn calculations for each. Nonparametric confidence intervals were then calculated for each candidate route to determine which choice or choices of route provided the minimal fuel burn. This new method was able to predict the most fuel-efficient flight route in approximately 30% of the routes examined while both the current AMC method and Homan (2014) were not able to identify the optimal route in any of the test cases. In this paper, the Boone (2018) method is first expanded by the inclusion of Bayesian prior estimates of fuel burn in confidence interval calculations for all candidate routes based on the previous month’s GEFS runs. The method is further expanded by incorporating conditional probabilities based also based on the prior month’s GEFS output. Example routes were tested and improvements in forecast skill are presented.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner